This directory contains AzureML pipelines to run various datasets through a given Azure AI endpoints, and assess the results.
The LLM prompting is done using the guidance
package.
It is provided as an 'extra' and was not used to generated the reported results.
components
This directory contains the Python components which are used in the AzureML pipelinesenvironments
This directory contains the definition of the AzureML environment shared by the various componentspipelines
This directory contains the code required to submit the pipelinesrequirements.txt
A standardpip
file which will install the necessary packages for the pipeline submission to work
Furthermore, the actual guidance
programs are in the top level guidance_programs
directory in this repository.
In order to submit a pipeline, you will need to give various pieces of information to the submission script (e.g. the AzureML workspace information).
Look in the pipelines/configs
directory, and you will see a number of *_template.yaml
files.
You will need to make copies without the '_template' suffix, and fill out the contents.
For exmaple, the aml_config_template.yaml
needs to be copied to aml_config.yaml
(in the same directory) and filled out with appropriate information.
The pipeline submission scripts all have names prefixed with submit_
.
To run one:
python ./submit_mmlu_zeroshot.py -cn zeroshot_config
where zeroshot_config
means the zeroshot_config.yaml
file in the configs
directory.